Abstract

An important part of the picture of mind developing with the recently resurgent connectionist paradigm in cognitive science is a radical view of mental representation. On this view, knowledge is represented in distributed form in patterns of synaptic strengths and neural activations in the brain. But what exactly is distribution? Surprisingly, this question has never been given a general and systematic answer, yet the issue is at the heart of many current debates, such as the proper characterization and relative merits of connectionist models. ;Researchers claim to have found distributed representations in diverse places; unfortunately, their characterizations are equally diverse. This thesis takes up and develops the common intuition that distributed representations form a natural genus. The essence of distribution turns out to be semantic superposition, i.e., multiple contents represented over the same space. Distributed schemes are therefore radically distinct from more familiar styles in that they generate a single point in the space of representations functioning as a representation of many contents simultaneously. Superposition is given a rigorous definition, leading in turn to a precise mathematical framework which encompasses a wide range of paradigm cases of distributed representation, revealing both their deep similarities and their particular differences. ;With this precise characterization in hand, it is demonstrated that distributed schemes are essentially incompatible with both symbolic and imagistic modes as those are typically defined in cognitive science. Since this is a controversial result, some common sources of confusion are discussed, including the widespread tendency to assume that representations bear certain properties of their contents. The thesis closes by considering two broad kinds of reasons for supposing distribution to be the form of mental representation. First, there is plentiful neuroscientific evidence of distribution in the brain; and second, a broad reconceptualization of our "cognitive environment" shows the inherent plausibility of distribution at a purely psychological level